Synthesize, Prompt and Transfer: Zero-shot Conversational Question Generation with Pre-trained Language Model

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8989–9010 July 9-14, 2023 ©2023 Association for Computational Linguistics Synthesize, Prompt and Transfer: Zero-shot Conversational Question Generation with Pre-trained Language Model Hongwei Zeng1,2, Bifan Wei2,3∗, Jun Liu1,2, Weiping Fu1,2 1Shaanxi Provincial Key Laboratory of Big Data Knowledge Engineering, School of Computer Science and Technology, Xi’an Jiaotong University, China 2National Engineering Lab for Big Data Analytics, Xi’an Jiaotong University, China 3School of Continuing Education, Xi’an Jiaotong University, China hongwei.zeng@foxmail.com, {weibifan@, liukeen@, fuweiping@stu.}xjtu.edu.cn Abstract Conversational question generation aims to gen- erate questions that depend on both context and conversation history. Conventional works utilizing deep learning have shown promis- ing results, but heavily rely on the availabil- ity of large-scale annotated conversations. In this paper, we introduce a more realistic and less explored setting, Zero-shot Conversational Question Generation (ZeroCQG), which re- quires no human-labeled conversations for training. To solve ZeroCQG, we propose a multi-stage knowledge transfer framework, Synthesize, Prompt and trAnsfer with pRe- Trained lAnguage model (SPARTA) to effec- tively leverage knowledge from single-turn question generation instances. To validate the zero-shot performance of SPARTA, we conduct extensive experiments on three conversational datasets: CoQA, QuAC, and DoQA by transfer- ring knowledge from three single-turn datasets: MS MARCO, NewsQA, and SQuAD. The ex- perimental results demonstrate the superior per- formance of our method. Specifically, SPARTA has achieved 14.81 BLEU-4 (88.2% absolute improvement compared to T5) in CoQA with knowledge transferred from SQuAD. 1 Introduction Question Generation (QG) aims to automatically generate questions from the given context and an- swer. It plays a vital role in knowledge testing (Heilman and Smith, 2010; Lindberg et al., 2013; Ghanem et al., 2022) and information seeking (Shum et al., 2018; Rosset et al., 2020; Zamani et al., 2020) by creating quiz questions and span- ning question suggestions, respectively. Most exist- ing QG research has usually focused on generating single-turn questions, which are formalized as inde- pendent interactions (Zhou et al., 2017; Zhao et al., 2018; Tuan et al., 2020). However, it is a more natural way to achieve complex information need ∗ Corresponding author Conversational Question Generation Context: Friedrich, pausing at Gross-Nossen, and per- haps a little surprised to find no Loudon meddling with him, pushes out, first one party and then another, Dal- wig, Bulow, towards Landshut Hill-Country, to threaten Loudon’s Bohemian roads;–who, singular to say, do not hear the least word of Loudon thereabouts. Conversation History: Q1: Who paused at Gross-Nossen? A1: Friedrich Q2: What was he caught off guard about? A2: No Loudon meddling with him Answer: Dalwig and Bulow Question: What parties did he push out? Table 1: A conversational QG instance in CoQA dataset (Reddy et al., 2019). through conversations involving a series of inter- connected questions (Reddy et al., 2019). Different from single-turn QG, the task of conversational QG (Gao et al., 2019) aims to generate questions which depend on both context and conversation history. Recent conversational QG models (Gao et al., 2019; Pan et al., …

Content truncated to 500 words